机器人策略Robot Policy

策略是观测到动作的映射：给定当前状态或观测，策略决定下一步执行什么动作，可以是确定性规则，也可以是输出动作分布的随机策略。

机器人策略通常由学习得到——模仿学习从示范拟合它，强化学习按奖励优化它；训练完成后，部署到机器人上的就是这个映射本身。

策略与环境构成闭环：动作改变环境，新观测又喂回策略。机器人行为的连贯或失控，都要从这个闭环里找原因，而不是归结为某一条单步指令。

A policy is a mapping from observations to actions: given the current state or observation, it decides the next action, either as a deterministic rule or a stochastic policy outputting an action distribution.

Robot policies are usually learned: imitation learning fits them from demonstrations, reinforcement learning optimises them by reward; after training, the deployed artefact is this mapping itself.

Policy and environment form a closed loop: actions change the environment, and new observations feed back into the policy. Coherent or runaway behaviour should be traced through this loop, not blamed on one single-step command.